# Abstracting parameters of dynamic model from output time series

I am unable to identify general temrs or specific source of information for the below proposed problem. I would appreciate if the community can guide me to journal articles/books and keywords to look for in literature.

Problem:

There is a non-linear dynamic system taking input and producing 1D time series as output. I would like to use NN to find parameters of the dynamic system, according to the time series output. That is, mapping the features of the time series (after transformation, likely Fourtier Transform or Wavelet) to the parameters governing the dynamics of the system.

Research so far:

I have found a few journal papers mostly processing sounds of rolling bearings or hearbeat but only for error/failure classification.

1. Rolling Bearing Fault Diagnosis Based on STFT-Deep Learning and SOund Signals
2. Deep Learning Enabled Fault Diagnosis Using Time-Frequency Image Analysis of Rolling Element Bearings
3. Deep Learning Based Approach for Bearing Fault Diagnosis
4. Detecting atrila fibrillation be deep convolutional neural networks

(the above are classification problems, my problem is about parameter identification)

Reason to address this on StackExchange:

I think I am missing overview about the topic (identification of dynamic systems using NN), because I am not able to reach more profound information. Also, I think that NN would be more beneficial to my current application than lets say optimization by evolutionary algorithms, threfore I am specifically asking for NN.

• Welcome to ai.se....What parameters are you trying to find out exactly? Is it discrete in nature or continuous...I think you did not formulate the problem properly...You need some more info...And if your primary aim is fault detection it his highly unlikely a nn will help yoi – DuttaA Apr 25 '18 at 11:33
• @DuttaA For example, if I have LTI, then I want to recover (at least partialy) the matrix A, as in y = A.x + b.u (x - state vector, u - input). By the way, why is fault detection a problem ? How about those journal papers I have provided (?), they detect/classify faults from the behavior/output of the system. – new_stacker Apr 25 '18 at 13:10
• Fault detection is a different subclass of problem assuming it is a skewed class..That is number of faults<<number of corrects – DuttaA Apr 25 '18 at 13:34
• @DuttaA So any suggestion for my question ? – new_stacker Apr 25 '18 at 14:18
• Nopes..I have very less understand of how things work out in the frequency/time domain Especially the imaginary part..But if you want to get an answer you should provide the links of the works instead of just names and can also post the question on electricalE.se, signal processing.se and crossvalidated.se – DuttaA Apr 25 '18 at 15:18